The Mechanical Manipulation of Randomly Oriented Parts

C onsider the fine coordination between the eye and the hand of a young child who picks a cookie out of a jar. Although the cookies are roughly uniform in size and shape, the pile of cookies at the top of the jar is a jumble of visual cues, a rugged topography from which the child must extract enough information to determine what part of the visual or tactile field can be ascribed to the single, target cookie. As the child learns to take a cookie without crushing or breaking the ones around it, the child comes to realize that not every 4 orientation of the hand can be successful. For example, seizing the edge of the cookie between thumb and forefinger works only if the center of the cookie is 1 on or near the line connecting the opposing points of pressure. A much more reliable strategy is to determine the attitude, or orientation, of the cookie visually and then turn the hand to one of the positions best suited for picking it up. Finally, having grasped the cookie in one attitude or another, the child must transform the spatial coordinates of the cookie that pertain to the hand into the coordinates that pertain to the mouth. Until recently such a complex set of coordinated actions was beyond the capability of mechanization that seeks to replicate some of the functions of factory workers. The robot now working in the factory is fundamentally a playback machine for motions in space. T o carry out a task the robot must f i s t be “trained” by a person already skilled in the task. The “arm” of the robot is guided through a series of motions, and the sequence of robot configurations needed to follow the trainer is recorded on a tape or other memory devid-When the tape is played back, it directs the robot to execute the same sequence of motions. The ability of the robot to record spatial motion has been exploited by choreographers to make a permanent record of dance movements, but without notable success. Nevertheless, the playback robot has found a niche in the factory because many industrial tasks are so highly repetitive that they can be done as a sequence of fixed motions. Mechanical manipulators have therefore been applied to spot welding, machine loading, painting, deburring. seam welding, sealing and other tasks that are boring or hazardous. There is much factory work that cannot readily be adapted to a fixed routine of movement. In manual assembly, for example, it is common to have parts stored in trays or bins surrounding the work station. There the blind playback robot is virtually useless because it can tolerate very little uncertainty in the position of a part it must handle. An obvious solution to the problem is to avoid jumbling the parts together in the first place, or in other words to maintain a controlled orientation from the time they are made. There is a trend among manufacturers in favor of this solution: parts can be organized on carriers or attached to pallets on which they can be mechanically manipulated without the need for sensing. Nevertheless, the solution has its costs. The carriers or pallets must be designed and manufactured, often to close tolerances. Moreover, the pallets are usually heavy, they take up a large amount of space and they often have to be redesigned when the part they carry is modified. Indeed, the design of the part itself may have to be altered for the sake of automatic feeding. Suffice it to say there are many circumstances in which the volume of production has not presented enough economic incentive for the manufacturer to depart from more traditional, manual methods. We have now developed a computer system that can determine the position of a part with an arbitrary shape in a randomly arranged pile. The system requires only a few electronic images of the pile of parts. The images are mathematically transformed by the computer into a form that is readily compared with a mathematical model of the part stored in the computer memory. The mathematical model is rotated by the computer until it closely matches the attitude of the object to be grasped. The results are applied to direct a mechanical arm to pick up the part. Such a flexible sensing system may be able to substantially extend the range of applications of industrial robots.